ECNDT 2006 - Tu.2.7.

Response to API 1163 and Its Impact on

Pipeline Integrity Management
Munendra S TOMAR, Martin FINGERHUT; RTD Quality Services, USA

Abstract. Knowing the accuracy and reliability of ILI measurements is important

for determining optimum integrity rehabilitation project scope, re-inspection
interval and the risk associated with the pipeline section examined. The recent
(September 2005) introduction of API 1163 (title) highlights the importance of
these metrics. Qualifying these metrics requires verification measurements with
accuracies a level of magnitude higher than in-line-inspection. LasersureSM
provides a process which uses in-the-ditch Laser Profilometry measurements
which have an order of magnitude greater accuracy than the original ILI-MFL
measurements, therefore are well suited to qualifying and potentially improve ILI
results. The following paper outlines the LasersureSM process and the potential
impact of this process on pipeline integrity management process.

Introduction

In-Line Inspection (ILI)

ILI (MFL) is commonly used by operators to determine the location and extent of metal
loss (primarily) on their pipeline systems. MFL technology and its understanding have
improved by leaps and bounds in the last decade or so, and has proven to be a valuable
tool for operators. Despite the advances, significant uncertainty still exists regarding the
accuracy of the data obtained from the MFL tools. Various factors such as the cleanliness
of the pipe, chemical and physical properties of the steel, presence of localized residual
stresses, variation of wall thickness, etc. all have a detrimental effect on the accuracy that
can be expected from ILI.
ILI Tool vendors typically specify the accuracy and certainty associated with their
data. In some cases, such metrics are defined for different geometrical categories as by
the Pipeline Operator’s Forum and API 1163. In other cases, it is provided as an
aggregate metric. The accuracy, thus defined, is typically of the order of +/- 10% with a
certainty of 80% with respect to the depth measurement. In other words, the depth
measurements from the MFL data would typically have an error, no greater than 10% of
the wall thickness in 80% of the data points. This still leaves a huge amount of
uncertainty that an operator has to account for while making critical decisions.
This uncertainty in the data has direct impact on the risk level that an operator associates
with a pipeline system or the inspected section there-of. The economic impact on the
integrity management program(s) can be huge, as operators have to make conservative

1
estimates of the associated risk, subsequently driving up the number of digs in a
rehabilitation project and/or shortening the re-assessment interval.

LaserscanSM

LaserscanSM is a service provided by RTD wherein the LPITTM tool uses laser
profilometry for mapping the external surface of the pipe and thereby measuring with
millimeter level resolution the extent of external corrosion as present on the pipe. As of
now, it is a benchmark technology and has been proven to be very accurate and reliable
in the field. It is capable of providing a high definition 3-dimensional map of the pipe
surface in a digital format. Due to the nature of the dataset, the data from LaserscanSM is
ideally suited for detailed analysis and comparison.

API 1163

API 1163 (In-Line Inspection systems qualification standard) was introduced in

August 2005 as an umbrella document defining the manner in which the ILI data was to
be reported and qualified. While it does define the specifications for anomalies belonging
to different geometrical categories, it does not offer a robust process for comparing the
ILI data with the field measurements.
A metal loss defect, as measured in the field, usually comprises of multiple
indications (each of the same or different geometrical shape) combined together using a
pre-defined interaction criteria. Similarly, ILI data also undergoes a “clustering process”.
To classify the indications into different geometrical categories, one has to use the un-
clustered data.

Previous work

Several efforts have been made in the past to do correlation analysis between MFL
data and field measurements. While considerable insight was gained in the process, most
of these efforts were marred by a few common drawbacks. Some of these are as listed
below:
• The field data was usually collected using manual methods, and therefore, was
limited in definition and accuracy
• Manual measurements typically are prone to measurement and indexing error,
making it difficult to match with the corresponding ILI anomalies
• Field data resolution is generally low, limiting the ability to match data points by
geometrical shape.

Presented study

The present study is based on the premise that the MFL tool has a different level of
measuring accuracy for anomalies belonging to different geometries. For the purpose of
this study, the anomalies were classified into the following categories as defined by the
Pipeline Operators Forum:
• Extended/General
• Pitting

2
 • Pinholes
• Axial Grooves
• Axial Slots
• Circumferential Grooves
• Circumferential Slots
There are a few key characteristics that differentiate the study performed from the
previous works in this direction. They are summarized as follows:
• The data points were matched on an individual pit basis
• The ILI data used was boxed data (prior to the application of interaction rules)
• The analysis was performed on data sets belonging to each geometrical category
individually
• All data points were taken into consideration, including corrosion with very small
depth measurements (>= 0.3mm)
Also, some assumptions were made to apply the process:
• It was assumed that the error in the LPITTM measurements was negligible
• It was assumed that the measurements from the ILI data were reliable enough to
determine the geometrical category a particular indication belonged to.

Process

The process used in this study comprised the following steps:

1. Data Assimilation
2. Classification
3. Matching the data points
4. Data Extraction
5. Analysis
6. Validation

Data Assimilation:

The boxed data from a previous ILI (MFL) inspection was provided for this study
by a major North American pipeline operator. The inspection was performed over three
valve sections. In addition, LaserscanSM service had been utilized during the
rehabilitation project previously carried out by the operator. The data from all the
locations inspected using the LaserscanSM and the MFL data from the corresponding pipe
segments was located.

Classification

Each MFL data point (indication) was then classified into one of the seven
categories based on the ILI reported length and width using the criteria as defined in the
POF document.

3
Matching the data points

The LPITTM data and the MFL data was then overlaid using a graphical tool
designed for the purpose. Some key indications were used to determine if there was any
location error in the data sets and data points that overlaid well were documented.
Following is a snapshot of the utility used to match the data points (Figure 1). The blue
indications are LPITTM indications and the red ones are MFL indications.

LPIT
ILI

Figure 1: Data matching utility used to overlay the ILI and LPIT datasets
In cases where on indication in either datasets corresponded to multiple
indications on the other, the boundary conditions were used to aggregate the multiple
indications so as to get a one to one correlation.

Data Extraction

The data from the matching data points was then accumulated and verified to
ensure consistency. The parameters for consideration (Length and Depths) for these
indications were extracted from the pool of data and arranged in a comparable format.
Some of these matching data points were set aside as a control group at random for
validation purposes.

Analysis

Statistical analysis was performed on these data sets. To streamline the process,
and due to the limited amount of data available, the analysis was performed for two of the
geometrical categories, viz. General corrosion and pitting. Within these categories, both
parameters of interest (Depth and Length) were evaluated to determine the error and any
evident bias in the positive or negative direction. Figure 2 shows the potential impact an
error in length measurement can have.

4
 Interacting
Measured

Actual length

Non-interacting

Figure 2: Anomalies clustered due to error in length measurement

Validation

Once the bias (or relation) was determined, a correction factor was calculated and
applied to the MFL data points from the control data set. These corrected measurements
were then compared against the LPITTM measurements to determine if there was any
improvement in the accuracy. The results are as provided in the following sections.

Data Analysis

The data considered for the purpose of this study was taken from 6 different
excavation locations. Data extraction provided us with 138 indications belonging to the
pitting corrosion category and 59 indications belonging to the general corrosion category.
38 indications from the pitting corrosion data set and 20 indications from the general
corrosion data set were randomly set aside as the control group for validation purposes.
The LaserscanSM measurements were used as the reference data set to calculate the error
in depth as well as length measurements for each data point.
Different standard statistical techniques were utilized to measure the bias, if any, in the
ILI data using the Depth and Length as the parameters of interest. They are:
1. Paired t-test [1]
2. Wilcoxon signed ranks test [2]
3. Passing-Bablok method [3]
The two data sets, ILI and LaserscanSM, were treated as paired data, as it was assumed
that since both measurements are for the same anomaly, values in one data set would
change with the other [4][5][6]. As is shown in Figure 3, the bias plots [4] for both the
depth and length measurements show a clear bias. The depth measurements reported by
ILI seem to be almost consistently less than the corresponding LPITTM measurements for
both pitting and general corrosion.

5
a)
25
Difference between metho
20

15

10

0 Zero bias

-5

-10
5 25 45
0 20 40 60
Mean of Depth

b)
5

0 Zero bias
Difference between metho

-5

-10

-15

-20

-25

-30
0 10 20 30 0 20 40 60
Mean of Length

c)
30
Difference between metho

25
20
15
10
5
0 Zero bias

-5
-10
5 15 25 35 0 5 10 15
Mean of depth

6
d)
250
Difference between metho
200
150
100
50
0 Zero bias

-50
-100
-150
0 100 200 300
0 10 20 30
Mean of Length

Figure 3: Bias plots showing that a bias is present in both length and depth measurements. (a) and (b) are
Bias plots for Depth and Length respectively for Pitting Corrosion. (c) and (d) are bias plots for Depth and
Length for General corrosion category

The calculations for Depth

The error in depth measurement (LPITTM – ILI) seemed to follow the normal distribution.
This was verified using the kolmogorov-smirnov test at a significance level of 0.05
(Figure 4).
(a)

14

12

10
Frequenc

Coefficient P
Kolmogorov-
Smirnov 0.6261 > 0.15
Skewness 0.3370 0.3540
Kurtosis -0.3670 0.7007

7
(b)

45

40

35

30
Frequenc

25

20

15

10

Coefficient P
Kolmogorov-
Smirnov 0.6586 > 0.15
Skewness 0.2638 0.2653
Kurtosis 0.7096 0.1579

Figure 4: Results from Kolmogorov-Smirnov test demonstrating that the error does follow a normal
distribution. (a) and (b) are histograms for Error in Depth for Pitting and General Corrosion respectively.

Now, assuming normality, an estimation of the bias (assuming a constant bias)

was performed at a significance level of 0.05. Since the mean error was found to be
positive, the estimated bias was added to the depth measurements in the control group for
validation and the results compared to the uncorrected error. The results are as provided
in Table 1.
Similar calculations, when carried out on the control data for general corrosion,
revealed a similar phenomenon. The SSE for uncorrected measurements was found to be
1346, while the SSE for corrected measurements was found to be 744. The data, when
subjected to the non-parametric wilcoxon signed ranks test revealed results almost
identical to the results presented.
Further investigation of the data seemed to indicate that the magnitude of the error
seemed to be a function of the reported depth measurement itself. This trend is quite
evident in the following charts (Figure 5)

8
 Error VS ILI Depth (General) Error VS ILI depth (Pitting)
25
12
20 10
8
Erro r

15
6
4

E rro r
10
2
5 0

6
8
10
11
11
12
13
13
14
14
15
17
18
18
19
21
22
24
26
33
-2
0
-4
10
11
11
11
13
13
14
15
15
16
19
22
23
24
6
6
7
8
8

ILI- depth ILI-Depth

Figure 5: Plots showing that the magnitude of error could be a function of the measured value

This indicated that the bias in the data might be proportionate to the depth.
Therefore, the data was then subjected to the Passing-Bablok comparison test. The test
was carried out at a confidence level of 95%.The results from the tests are as follows
(Figure 6):
Identity line
55 Y =X 35
50
Pitting General
30
45

40 25
Depth - ILI Dep

35
20
ILI - D

30
15
25

20 10

15
5
10 y = 0.7833x - 0.3869 y = 0.7585x - 2.5196

5 0
5 15 25 35 45 55 5 15 25 35
LPIT - D
Depth - LPIT Depth

Figure 6: Results from Passing-Bablok regression reflecting the proportionate bias

Subsequently, the control data was corrected using the respective correction model. The
results showed marked improvement over the uncorrected data. The results are as
provided in Table 1.
One thing to be noted here is that while the sum squared error seems to be lower for the
non-parametric wilcoxon-bablok method, the absolute maximum error was found to be
more in the parametric case. In other words, using a constant bias would cause the
outliers to be even more pronounced.

The calculations for Length

The MFL tool reports depth as a relative measurement as a percentage of the wall
thickness. This normalizes the data and therefore, the scatter doesn’t seem as pronounced.
Since the length measurements are absolute (mm), the effect of outliers on the analysis is

9
quite pronounced. To reduce this, the analysis for length was carried out using percentage
length, defined as the ratio of the lengths reported by LPITTM versus that by ILI times
100. The same calculations, when carried out for length, yielded results as shown in
Table 2.
As is obvious from these results, the improvement in length is not nearly as
significant as for depth, although some improvement can be achieved. The reduced
improvement can partly be attributed to the fact that the deviation from true measurement
(error) is much larger for length than for depth.

Conclusion

The following key conclusions can be derived from this exercise:

• There is statistically significant bias in both depth and length measurements in the
MFL data.
• The bias is different for different geometrical categories. While the process, as
defined in this paper was completed on pitting and general corrosion, it can also
be applied to calculate the ILI tool performance for different geometries of
anomalies.
• While conventional (parametric) statistics does provide means for calculating
these biases, these techniques typically assume that the data belongs to a
particular distribution and therefore, when a sample does not conform to the given
distribution, the error in the outliers is magnified causing larger maximum
absolute errors.
• Non-parametric techniques seem to be more adequate as they are more resilient to
outliers in the data set.
• These correction factors can be applied to boxed data and the data clustered again
by applying the respective interaction criteria to provide more accurate data to the
operators, thereby reducing the overall uncertainty and subsequently, risk.
• Development and subsequent application of such a process can assist operators in
establishing qualified re-assessment intervals owing to higher certainty in the
data.
• Methods such as bootstrapping and/or Bayesian techniques should be explored to
determine if they can provide further improvement in results.

References

[1] Statistical methods for assessing agreements between two methods of clinical measurement. Bland
J.M., Altman D.G.
[2] Individual comparisons by ranking methods. Wilcoxon F. (1945)
[3] A general regression procedure for method transformation. Passing H., Bablok W
[4] Principles and procedures of exploratory data analysis. Behrens, J. T. (1997).
[5] NIST/SEMATECH e-Handbook of Statistical Methods,
http://www.itl.nist.gov/div898/handbook.
[6] Essentials of Statistical Inference (Cambridge Series in Statistical and Probabilistic Mathematics)
2005.

10
 Table 1: Pitting corrosion
Corrected Depth Corrected Length
Using Using Using Using
Uncorrecte constant Passing Uncorrecte constant Passing
d depth bias Bablock d Length bias Bablock
Mean
Error 5.78 1.93 0.22 -37.34 -0.55 -37.62
Standard
Deviation 5.00 5.00 5.46 68.36 50.05 64.67
Max Error 19.37 15.52 13.40 233.33 144.05 226.22
Sum
squared
error 2193.77 1066.87 1104.82 225889.27 92698.66 208516.21

Table 2: General corrosion

Corrected Depth Corrected Length
Using Using Using Using
Uncorrecte constant Passing Uncorrecte constant Passing
d depth bias Bablock d Length bias Bablock
Mean
Error 8.28 1.39 -1.27 -7.67 -5.16 4.74
Standard
Deviation 9.54 9.54 10.72 41.95 40.23 23.43
Max Error 27.78 20.89 20.96 112.00 104.55 39.44
Sum
squared
error 1345.57 745.81 934.02 14605.00 13189.95 4595.61

11